Git history vs AI exploration
Git tells us what changed. It doesn't tell us what was tried, learned, rejected, or abandoned. Where should that knowledge live for AI agents?
The more we use AI agents, the less valuable commit history becomes as the sole source of truth.
Git tells us what changed.
It doesn’t tell us what was tried, learned, rejected, or abandoned.
A common scenario:
I create a branch to explore an idea.
exp/payment-retry-uiI spend a few days experimenting, get pulled into higher-priority work, and move on.
Six months later:
“Didn’t we already explore this?”
The code may still exist somewhere in Git history. The branch may still exist. There may even be a few WIP commits.
But how does an AI agent discover:
- Why the experiment was started?
- What conclusions were reached?
- What worked?
- What didn’t?
- Why it was abandoned?
- Whether it’s worth revisiting?
Possible approaches I’ve seen:
- Draft PRs as living experiment logs
- GitHub/GitLab Issues linked to branches
- Annotated tags on significant exploration commits
- Branch-local markdown notes
- External engineering worklogs
- AI-native memory tools that track experiments independently of the repository
Git is great at storing code history. It’s surprisingly bad at storing exploratory knowledge.
How should AI agents discover abandoned experiments from six months ago?
Where do you think this knowledge should live?